JP3232701B2 - Audio coding method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech coding method suitable for obtaining high-quality synthesized speech at a low bit rate, and more particularly to a speech coding method applicable to bit rates of 4 kbps or less with a relatively small processing amount. According to.

[0002]

2. Description of the Related Art Recently, a speech coding system incorporating an "analysis by synthesis" technique for evaluating a weighted error between a synthesized speech and an original speech and determining a coding parameter so as to minimize the error has been proposed. Has succeeded in obtaining relatively good voice quality even at low bit rates. A typical example is a code-driven linear predictive coding (CELP) scheme (for example, MR Schroeder and BS Atal: "Code-ex
cited linear prediction (CELP) ", Proc. ICASSP 85 (1
985.3)) and achieves practical voice quality at 4.8 kbps. Also, a number of improved CELP schemes have been proposed, such as a vector sum driven linear predictive coding (VSELP) scheme (for example, IA Gersonand MA).
Jasiuk: "Vector sum excited linear prediction (VSE
LP) speechcoding at 8kbps ", Proc. ICASSP 90 (1990.
4)) is superior in terms of processing amount, memory capacity, and bit error resistance.

[0003] On the other hand, digitalization of mobile radio communication is in full swing, and from the viewpoint of effective use of frequency, development of a voice coding system with a lower bit rate (4 kbps or less) is desired. If the bit rate of CELP or VSELP is simply reduced, quality degradation is increased and there is a limit. This is because the long-term prediction accuracy by the adaptive codebook search is reduced, and the reproducibility of the periodic component is reduced. As a result, the sense of noise in the decoded speech is increased. Therefore, a method has been proposed in which a pulsed sound source is introduced in addition to the conventional statistical sound source (noise source) to improve the reproducibility of the periodicity.

[0004] As such a system, for voiced sounds, a single pulse whose phase and amplitude are controlled, and for unvoiced sounds, CELP is used.
“SPE-CELP” method using WWW (Granzow and
BSAtal: "High-quality digital speech at 4 kb /
s ", Proc. GLOBECOM 90 (1990.12)) or" Pulse / Noise Selective CEL, which switches between periodic pulse and noise.
P "method (Yoshida et al. 2:" Application of pulse excitation search to low bit rate CELP coding ", IEICE Tech.
68 (1991.10), or Tanaka and Itakura: "C
Improvement of Speech Quality by Introducing Pulse Sound Source in ELP Speech Coding Method ", IEICE Technical Report EA92-24 (199)
2.5)).

[0005]

The speech coding method using the above-mentioned pulse sound source can improve the reproducibility of the periodic component even if the bit rate is reduced as compared with the conventional method, but has the following problems. .

[0006] Since the "SPE-CELP" system uses only one pulse per pitch period, its position and amplitude have a great influence on voice quality. The method of determining the pulse position is quite complicated, and has a problem in robustness to an input audio signal. There is also a report that coded speech may be buzzer-like.

[0007] On the other hand, the "pulse / noise selection type CELP" system evaluates an error when a pulse sound source and a noise sound source are separately used, selects a sound source having a smaller error, and determines voiced / unvoiced input voice. To select the sound source to use. In these methods, since long-term prediction (adaptive codebook search) is used together, the pulse sound source has a strong meaning to complement the long-term prediction vector. However, in the above document, since the pulse interval is limited to the long-term prediction lag or the pitch period, there is a problem that sufficient voice quality is not obtained.

[0008] Further, since both the "SPE-CELP" system and the "pulse / noise selection type CELP" system switch between a pulse sound source and a noise sound source, a change in timbre due to the sound source switching occurs in the coded speech. There is also a problem of (discontinuity).

A first object of the present invention is to provide a coding system in which the voice quality is less deteriorated even when the bit rate is reduced, and the change in timbre is not conspicuous. Further, a second object of the present invention is to realize the first object with a relatively low processing amount.

[0010]

In order to achieve the above object, the present invention has the following means in place of the statistical codebook and its search means in the structure of a normal CELP. (1) pulse information codebook, (2) pulse generation means, and (3) pulse sound source search means.

[0011]

The operation of the representative configuration of the present invention will be described.

The speech input to the speech encoder is first divided into frames and subframes. In the short-term prediction analysis unit, a spectrum parameter (short-term prediction coefficient) is extracted and quantized for each frame. Next, as preparation for evaluating the hearing weighting error, the input speech is subjected to hearing weighting. Also, a zero signal is input to the weighting synthesis filter, a zero input response is obtained, and the response is subtracted from the weighted input signal. This is to remove a past effect that depends on the internal state of the synthesis filter. Further, the impulse response of the weighting synthesis filter is calculated.

Next, in the long-term prediction analysis unit, an optimum long-term prediction lag and gain are obtained from the adaptive codebook in subframe units. From the signal obtained by subtracting the zero input response from the weighted input signal, a signal is generated by subtracting a weighted long-term prediction vector further multiplied by a gain, and is input to the pulse sound source search unit.

In searching for a pulse sound source, first, information on a pulse interval and a leading pulse position is read from a pulse information codebook, and a pulse generator generates a pulse train. This pulse train is regarded as a sound source vector, and weighting is performed by convolution of the impulse response of the weighting synthesis filter. A weighting error is sequentially evaluated for these weighting vectors, and a pulse excitation vector and a gain that minimize the error are determined.

The gain quantization unit simultaneously optimizes and quantizes the long-term prediction vector and the gain of the pulse sound source.

The spectral parameter, gain quantization code, long-term prediction lag, and index of the selected pulse excitation code vector determined as described above are transmitted to the decoder as transmission parameters.

In the decoder, a driving sound source is calculated from the transmission parameters, and is input to a synthesis filter using short-term prediction coefficients as filter coefficients, thereby obtaining decoded speech.

[0018]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of a speech encoding unit according to an embodiment of the present invention, and FIG. 2 is a block diagram of a speech decoding unit.

The present invention provides code driven linear prediction (CELP).
Since it is based on the speech coding method, the principle of the CELP method will be described first before describing a specific embodiment. FIG. 3 is a diagram illustrating the principle of determining the drive excitation in the CELP encoding unit. In the figure, a long-term prediction vector 110 which is the output of the adaptive codebook 108 as a component representing the periodicity of the sound source, and a code vector 111 which is an output of the statistical codebook 109 as a component (randomness, noise) other than the periodicity. Is multiplied by the respective gains 112 and 113, and the sum is used as the driving sound source.

A search for a code book for obtaining an optimal driving sound source is performed as follows. In general, it is sufficient that a driving sound source that is obtained by inputting a driving sound source to a synthesis filter matches the original sound (input sound), but in practice there is some error (quantization distortion). Therefore, it suffices to determine the drive sound source so as to minimize this error, but it is known that human auditory characteristics do not always correspond to the error amount and the subjective quality of speech. Therefore, it is common to use an error weighted so as to improve the correspondence with the auditory characteristics. The auditory weighting is described in the following document, for example. BS Atal
and JR Remde: "A new model of LPC excitation f
or producing natural-sounding speech at low bit ra
tes ", Proc. ICASSP 82 (1982.5).

To evaluate this auditory weighting error,
The driving sound source 114 is input to the weighted synthesis filter 105 to obtain a weighted synthesized voice 116. The input voice 101 also obtains a weighted input voice 115 through the auditory weighting filter 104, and obtains a weighted error waveform 117 by taking the difference from the weighted synthesized voice 116. Note that the filter coefficients of the auditory weighting filter 104 and the weighting synthesis filter 105 are obtained by inputting the input speech 101 in advance by LPC (linear prediction).
It is determined by the LPC parameters 103 obtained by input to the analyzer.

The weighted error waveform 117 is calculated by the squared error calculator 118 over the error evaluation section to obtain a weighted squared error 119. As described above, since the driving sound source is the weighted sum of the long-term prediction vector and the statistical code vector, the determination of the driving sound source results in the determination of a code vector index that determines which code vector is selected from each codebook. That is, the weighted square error 119 is calculated by sequentially changing the long-term prediction lag 106 and the code vector index 107, and the error minimizing unit 120 may select the one that minimizes the weighting error. Such a drive sound source determination method is called an “analysis by synthesis” method.

When the optimal driving sound source is determined in this manner, the multiplexing section 121 multiplexes the long-term prediction lag 106, the codebook index 107, the gains 112 and 113, and the LPC parameters 103 as transmission parameters, and I do. Also, the driving sound source 114 at this time
Is used to update the state of adaptive codebook 108.

If the above-mentioned "analysis by synthesis" method is to be executed faithfully, that is, if the long-term prediction lag and the index of the statistical code vector are simultaneously optimized while evaluating the weighting error each time, an enormous amount of processing is required. . Therefore, a technique such as sequential optimization is actually used.

On the other hand, in the processing in the decoder, first, the received data 222 is separated into various parameters by the demultiplexer 221. The adaptive code book 208 is searched based on the long-term prediction lag 206, and a long-term prediction vector 210 is output.
Further, the statistical code book is searched based on the code book index 207, and the sound source vector 211 is output. The long-term prediction vector 210 and the sound source vector 211 are multiplied by the respective gains 212 and 213, and the added signal is added to the driving sound source 214.
Is input to the synthesis filter 230. The filter coefficient of the synthesis filter is determined by the LPC parameter 203. The post filter 231 is not essential, but is often used to improve the subjective quality of the synthesized speech, and its output is the output speech 232.

FIG. 1 is a block diagram of a speech encoder according to an embodiment of the present invention, and FIG. 2 is a block diagram of a speech decoder. Hereinafter, an outline of the operation of the present embodiment will be described.

The audio encoder receives a digital audio signal 11 that has been A / D converted at a predetermined sampling frequency (usually 8 kHz).

A short-term prediction analyzer (LPC analyzer) 12 reads out audio data 11 having an analysis frame length and outputs a short-term prediction coefficient 13. The frame length is, for example, 40 ms (3
20 samples).

The short-term prediction coefficient 13 is quantized by a short-term prediction coefficient quantizer 14. The quantization code is output as a short-term prediction coefficient quantization index 18 as a transmission parameter. Further, the quantized value 17 of the short-term prediction coefficient is referred to in the processing of the next and subsequent stages.

Further, the input voice 11 is weighted by the auditory weighting filter 19, and a weighted voice 20 is obtained. On the other hand, a signal (zero input) 22 having a value of 0 for the frame length is input to the weighting synthesis filter 21 to obtain a zero input response 23. This is subtracted from the weighted input speech 20,
The weighted input speech 24 is obtained in which the influence of the past internal state of the weighting synthesis filter has been removed. Further, the impulse response 29 of the weighting synthesis filter is also obtained.

Since the long-term prediction analysis is performed by searching the adaptive codebook for each subframe, it is hereinafter referred to as an adaptive codebook search. Here, the subframe length is, for example, about 8 ms (64 samples). The adaptive codebook searcher 25 extracts a long-term prediction lag, which is a parameter representing the periodicity of speech, and outputs a long-term prediction lag index 31 and a long-term prediction vector 48.

The pulse generator 35 reads the pulse interval and the leading pulse position from the pulse information codebook 33, and generates a pulse train 36 based on the information 34. The pulse sound source search unit 32 regards the pulse train 36 as a sound source vector, and determines the impulse response 29 of the weighting synthesis filter.
Is weighted by convolution of From the signal 24 obtained by subtracting the quiescent response 23 from the weighted input signal 20,
Further, an optimum pulse sound source 49 is searched for a signal obtained by subtracting the weighted long-term prediction vector 30 multiplied by the gain. The index 38 of the pulse information codebook 33 corresponding to the optimal pulse sound source 49 is output.

The gain optimizer / quantizer 39 calculates and quantizes the optimum values of the long-term prediction vector 48 and the gain of the pulse excitation vector 49. Quantization code 4 at that time
4 is output.

The short-term prediction coefficients and gain quantization codes 18 and 44, the long-term prediction lag index 31 and the selected pulse information codebook index 38 obtained as described above are transmitted to the speech decoder as transmission parameters. Is done.

In the speech decoder, the index 54 of the long-term prediction lag
From the adaptive codebook 60 using the
1 is read out, and the pulse information codebook index 5
5, information 63 on the pulse interval and the leading pulse position is read from the pulse information codebook 62, and a pulse generator 64 generates a pulse sound source 65. The gains 58 and 59 are reproduced from the gain codebook 57 using the gain codebook index 53. Each of the code vectors 61 and 62 is multiplied by each of the gains 58 and 59 and added to generate a driving excitation vector 68.

By inputting the driving sound source 68 to the synthesis filter 71, a synthesized speech 72 is obtained. The short-term prediction parameters read from the short-term prediction parameter quantization code book 69 based on the short-term prediction parameter quantization index 56 are used as the filter coefficients of the synthesis filter 71. Finally, for the purpose of improving subjective sound quality,
The synthesized speech 72 is input to the adaptive post filter 73, and a final decoded speech 74 is obtained.

The decoded sound (digital signal) is DA-converted, converted to analog sound, and output.

Having described the outline of the present embodiment, the detailed functions of the main parts will be described next.

The short-term prediction analyzer (LPC analyzer) 12
A short-term prediction coefficient 13 representing a speech spectral envelope is extracted from the speech data 11 for each frame. Short-term forecast coefficient 1
3 is most commonly a linear prediction coefficient, but an equivalent parameter derived therefrom, the partial autocorrelation coefficient (PA
It is easily converted to RCOR coefficients, reflection coefficients) and line spectrum pairs (LSP parameters).

As a method of deriving the linear prediction coefficient, Dur
bin-Levinson's iterative method (by Saito and Nakata,
"Basics of speech information processing", introduced in Ohmsha, 1981) is common.
In addition to the above, the FLAT algorithm (established by the Radio System Development Center, "Digital Car Phone System Standard RCR STD-27" (hereinafter "RCR Standard"
Abbreviations) and the LeRoux method (Saito,
Nakata, described in the above-mentioned book). Also,
A method of converting linear prediction coefficients into LSP parameters is also described in the above-mentioned book by Saito and Nakata.

In this embodiment, the linear prediction coefficients 13 are converted into LSP parameters, then vector-quantized by the quantizer 14 and converted into quantized values 17 (code vectors 16 are sequentially read from the LSP codebook 15). However, the one with the smallest error is the quantization value.) LSP
It is known that parameters have better quantization characteristics than direct quantization of linear prediction coefficients (spectral distortion is small even if quantized with the same number of bits). The quantization method is
Depending on the allowable number of bits, scalar quantization, multi-stage vector quantization, vector / scalar quantization, or the like may be used. The quantization index 18 is output as a transmission parameter.

Next, the preprocessing for calculating the auditory weighting error will be described. In order to calculate the weighting error, first, the input speech 11 is weighted by the auditory weighting filter 19 to obtain the weighted speech 20. The weighting filter 19 is a short-term prediction coefficient (or equivalent parameter)
, And its specific format is as follows.

[0043]

(Equation 1)

Here, α _i is a filter coefficient (linear prediction coefficient), Np is a filter order, for example, Np = 10, and λ is a weighting parameter, usually λ = 0.8.

In general, the output of the synthesis filter is affected by the past state. Here, the influence of the past synthesis filter is removed from the weighted speech 20 in advance to reduce the amount of calculation. That is, data having a value of 0 (zero input 22) corresponding to the frame length is output to the weighting synthesis filter 21.
, A zero input response 23 is calculated, and the weighted speech 20
From the weighted speech 24 from which the past effects have been removed.
Get. The transfer function of the weighting synthesis filter 21 used here is as follows.

[0046]

(Equation 2)

This synthesis filter 21 is different from the decoding-side synthesis filter in that it includes a weighting parameter λ. The impulse response 29 of the weighting synthesis filter 21 is also obtained at the same time. At this time, (Equation 2)
Is used as the quantization value 17 of the linear prediction parameter.

As described earlier, the long-term prediction analysis is regarded as an adaptive codebook search, and the long-term prediction lag (index of the adaptive codebook) is selected by minimizing the perceptual weighting error between the synthesized waveform and the original speech. . Here, a case where the adaptive codebook search and the pulse sound source search are sequentially performed will be described. That is, the optimal long-term prediction lag index 31 is determined on the assumption that no pulse sound source is used.

Next, the adaptive code book search section 25 will be described. The weighting synthesis of the code vector 27 read from the adaptive codebook 26 corresponding to the long-term prediction lag to be searched is realized by convolution with the impulse response 29 of the weighting synthesis filter. The synthesized output (weighted long-term prediction vector) 30 obtained in this manner is called a zero-state response because it does not depend on the past state of the synthesis filter. The long-term prediction vector 30 for each lag in the search range is calculated, the correlation with the weighted speech 24 from which the influence of the past is removed is calculated, the (optimum) long-term prediction vector 48 giving the maximum value of the correlation, A long-term prediction lag index 31 obtained by quantizing the prediction lag is output. For details of the long-term prediction analysis method and the method for reducing the amount of calculation, refer to the aforementioned RCR standard.

Next, generation of the pulse excitation vector will be described.

The present invention is characterized in that a pulsed sound source is used instead of the conventional CELP statistical sound source. The pulse sound source is basically a part (subframe length) of a periodic pulse train. However, as shown in FIG. 5, the head pulse position can be taken from the first sample to the last sample of the subframe regardless of the pulse interval. This is because the characteristic of the rising edge of speech, which cannot be covered by the long-term prediction vector, is reproduced by the pulse sound source due to the increase in the sub-frame length accompanying the lower bit rate. Further, the pulse interval is preferably set so as to substantially cover the fluctuation range of the pitch period of the human utterance, similarly to the search range of the long-term prediction lag. In this embodiment, the minimum pulse interval is Lmin = 21, and the maximum pulse interval is Lmax = 146.

The pulse information codebook 33 stores pulse intervals and leading pulse positions as shown in FIG. As can be seen from FIG. 5, when the pulse interval is L and the subframe length is N (N = 64 in this embodiment), L ≧
In the case of N, the number of pulses in the subframe is one. L
In the case of <N, the number is one or more depending on the leading pulse position. In the case of one line, since the overlap is made with the case of L ≧ N, the pulse interval and the leading pulse position are arranged in the pulse information codebook so that the pulse train does not overlap. That is, when L <N, the head pulse position is set to a range in which two or more pulses exist in the subframe.
For ≧ N, L = N is represented, and the leading pulse position is from 0 to N−1. In this embodiment, N = 64, Lmin
= 21, the type of non-overlapping pulse train is 1
010 patterns can be represented by 10 bits.

The pulse generator 35 generates a pulse as shown in FIG. 7 based on the pulse interval read from the pulse information codebook 33 and the information 34 on the leading pulse position. The amplitude of a pulse is 1, and the amplitude of a sample without a pulse is 0.

The above is a case where the pulse information vector 36 is generated by the pulse information codebook 33 and the pulse generator 35. However, it is of course possible to store all the pulse information vectors in the codebook. In this case, the pulse generation process can be omitted. On the other hand, the storage capacity of the codebook is N words, whereas the pulse information codebook 33 requires only two words of the pulse interval and the first pulse position per vector. become.

Next, search for a pulse sound source will be described.

First, the optimal long-term prediction vector 48 output as a result of the adaptive codebook search is represented by b _L (n), and its weighted signal (zero state response of b _L (n)) 30 is represented by b ′ _L (n ),
Let the gain be β. The weighted input speech 24 from which the influence of the past has been removed is defined as p (n). Where p '
Define (n).

[0057]

(Equation 3)

This represents a component obtained by subtracting the contribution of the long-term prediction vector from the ideal synthesized speech, and is a component to be covered by the pulse sound source.

Assuming that the generated pulse sound source is f _i (n) and the weighted synthesized speech is f ′ _i (n), the error E,

[0060]

(Equation 4)

It is sufficient to find f ′ _i (n) that minimizes f ′ _i (n). Here, γ _i represents a gain, and i represents an index (index) of the pulse information codebook.

If (Equation 4) is partially differentiated with γ and set to 0, γ _i that minimizes the error E is

[0063]

(Equation 5)

Where E is

[0065]

(Equation 6)

Is obtained. Here, the first term on the right side of (Equation 6) has a positive constant value regardless of f ′ _i (n), so f ′ _i (n) that maximizes the second term on the right side, that is, the pulse sound source f _i ( n).

The above processing is basically performed by the conventional CELP
This is the same as the statistical codebook search in, and is a large part of the processing amount. In the present invention, the amount of search processing is significantly reduced by using an impulse response whose order is truncated using the characteristics of a pulse sound source.

In general, when speech is synthesized by convolution of an impulse response, truncation of the order of the impulse response causes an error. However, as shown in FIG. 8, the impulse response of the weighted synthesis filter represented by (Equation 2) has a steeper attenuation than the impulse response without weighting.
The effect of order truncation is small. Set the truncation order to 20 (2.
If it is set to about 5 ms), in most cases, the effect of the discontinuation can be ignored. Therefore, in the present invention, the truncation order is set to Lmin (21 samples) which is the minimum pulse interval of the pulse sound source.

Here, C _i and G _i are defined as follows.

[0070]

(Equation 7)

C _i is the cross-correlation between p ′ (n) and f ′ _i (n),
The 'because it is the power of _i (n), would otherwise f' G _i is f _i
Each time (n) changes (every time the index i is updated), it is necessary to recalculate. On the other hand, p ′ (n) (0 ≦ n ≦ N−1, N
Is the number of subframe samples) and the impulse response h (n)
Is constant in a certain subframe. Where the order is Lm
The impulse response censored at in is represented by h ′ (n) (0 ≦ n ≦ L
min) and a _j (0 ≦ j ≦ N−1) represented by the following equation:
Is calculated in advance.

[0072]

(Equation 8)

A _j indicates the cross-correlation between the position of h ′ (n) and the portion of p ′ (n) corresponding to h ′ (n) when the position of h ′ (n) is shifted by one sample, as shown in FIG. I have.

Since h '(n) is spelled out at Lmin, there is no overlap between pulses for any pulse sound source to be searched. Accordingly, the a _j To obtain the C _i, for example, as shown in FIG. 10, when the pulse position of the pulse excitation f _i (n) is that it was P1, P2, P3, which is previously calculated in equation (7) Of these, the sum of a _P1 , a _P2 , and a _P3 may be calculated. Therefore,
Since the convolution calculation of the impulse response to be performed every time f ′ _i (n) changes is replaced with the sum of partial cross-correlations calculated once in the subframe in advance, the processing amount can be significantly reduced. became.

[0075] can also be applied to G _i of the same techniques (number 7). That is, g _j defined in advance by the following equation
Is calculated.

[0076]

(Equation 9)

Note that, as shown in (Equation 9), 0 ≦ j ≦ N
In the case of −Lmin, the value of g _j is constant, so that only g ₀ needs to be calculated. As with the calculation of the G _i also of C _i,
This can be realized by _obtaining the sum of g _j corresponding to the pulse position of f _i (n).

When the optimum pulse sound source f _i (n) (maximizing C _i ² / G _i ) is determined in this way, the impulse response h (n) without order truncation is used.
f _i strict weighting signal f _'i of (n) (n) previously calculated.

In the conventional method using the pulse code book (the above-mentioned document by Yoshida et al. And the document by Tanaka et al.), The pulse interval is a long-term prediction lag or a pitch period obtained by extracting a pitch. Therefore, when a pulse sound source is used in a portion where the periodicity of the input voice is low, the sound quality is deteriorated. According to the present invention, since all the pulse sound sources of the possible combinations are searched, the long-term prediction vector is complemented even in such a portion, and good sound quality can be obtained. As a result, there is no need to use a noise sound source, and problems such as a change in timbre due to sound source switching can be avoided.

The processing at the final stage in the speech encoder is optimization of gain and quantization. In the gain optimization / quantizer 39,
A strictly weighted long-term prediction vector b ′ _L (n) 30 (determined by convolution of the impulse response without order truncation) and a pulse source vector f ′ _i (n) 37, and
Weighted input speech p (n) 24 from which past effects have been removed
Is entered. Here, assuming that the gains are β and γ again, β and γ are determined so as to minimize the weighting error E in the following equation.

[0081]

(Equation 10)

More specifically, the equation (10) is partially differentiated by β and γ to solve a simultaneous equation that can be set to 0.

The quantization of the gain includes a method of directly scalar quantizing β and γ and a method of vector quantizing after transforming β and γ into another variable. In this embodiment, vector quantization is performed by the latter method. Refer to the RCR standard for the specific method.

Assuming that the quantization values of β and γ are β _q 42 and γ _q 43, respectively, the unweighted long-term prediction vector 48 and the pulse excitation vector 49 are multiplied, and the driving excitation 4
5 is produced. The driving sound source 45 is used for updating the adaptive codebook 26.

Next, returning to FIG. 2, the speech decoding section of this embodiment will be described.

The received data 51 is demultiplexed in the demultiplexer 52 into a short-term prediction parameter quantization index 56, a long-term prediction lag index 54, a pulse information codebook index 55, and a gain quantization index 53.

The first stage of the decoding process is to decode each parameter value. The short-term prediction parameter value 70 is decoded from the short-term prediction parameter quantization codebook 69 based on the short-term prediction parameter index 56. Similarly, the adaptive codebook 60 decodes the long-term prediction vector 61 based on the long-term prediction lag index 54. Gain codebook 5
7, the quantization gain 5 based on the gain quantization index 53.
Decode 8, 59. Pulse information codebook index 55
, The information 63 of the pulse interval and the leading pulse position is read from the pulse information codebook 62, and the pulse generator 64 generates a pulse excitation vector (pulse train) 65.
Is decoded.

The second stage of the decoding process is the generation of a driving sound source. Long-term prediction lag index 5 from adaptive codebook 60
4 are multiplied by the gains 58 and 59, respectively, and the driving sound source 68 is generated. The driving sound source 68 is input to the synthesis filter 71 and is also used for updating the state of the adaptive codebook 60.

The last stage of the decoding process is speech synthesis. The synthesis filter 71 uses the decoded short-term prediction parameter 70 as a filter coefficient and inputs a driving sound source 68 to synthesize and output a digital synthesized speech 72.
Furthermore, in order to enhance the subjective sound quality, the synthesis filter 71
Is passed through a post filter 73 to obtain a final digital synthesized voice 74 as the output. This is continuously sent to a DA converter via a buffer memory and converted into an analog synthesized voice.

The operations from the speech input to the encoding, decoding, and speech output according to the embodiment of the present invention have been described above. In the above description, the frame energy (power) of the speech has not been particularly mentioned. This is because the frame energy is reflected in the gain of the driving sound source. In consideration of the quantization of the gain, it is more advantageous to normalize the frame energy in advance in order to suppress the dynamic range of the gain. Since the frame energy can be easily obtained when calculating the linear prediction parameter, the frame energy is separately quantized and its index is transmitted. An example of bit allocation in such a case will be described below.

The sampling frequency is 8 kHz and the frame length is 4
0 ms (320 samples), subframe length 8 ms
(64 samples). It is assumed that the frame energy and the linear prediction parameter are updated on a frame basis, and the other parameters are updated on a subframe basis. Note that it is more effective to improve the quality of synthesized speech if the frame energy and the linear prediction parameter are interpolated and used in subframe units. Assuming that the quantization is performed by 27-bit multistage vector quantization, the quantization index of the linear prediction parameter is 27 bits. The frame energy is scalar quantized with 5 bits. Therefore, the number of transmission bits per frame is 3
Two bits.

The sub-frame unit parameter has a long-term prediction lag index of 7 bits, and the long-term prediction lag has a range of 19 samples (421 Hz) to 146 samples (5 samples).
5 Hz). The codebook size of the pulse information codebook is set to 10 bits (1010 code vector,
If 14 vectors are not used), the code vector index is 10 bits. The gain is converted into another parameter for the long-term prediction vector and the one for the statistical code vector, and then vector-quantized and represented by 8 bits.
Therefore, the number of transmission bits per subframe is 25 bits. From the above, the total bit rate is 3925
bps.

As described above, in the embodiment of the present invention, even if the bit rate is reduced, the reproducibility of the periodic component is improved, and high quality can be achieved. Also, by searching the sound source codebook by the combination of the impulse responses with the order censored,
The processing amount can be reduced as compared with the conventional CELP.

[0094]

According to the present invention, the reproducibility of the periodic component, which is a problem when the CELP encoder is reduced in bit rate, is improved, and a noise source is not required.
A speech encoder with good speech quality can be provided even at a bit rate of s or less. Further, since the amount of processing for searching for a pulse sound source can be reduced, a speech encoder with a low processing amount can be provided.

[0095]

[Brief description of the drawings]

FIG. 1 is a block diagram of an encoding unit according to a first embodiment of the present invention.

FIG. 2 is a block diagram of a decoding unit according to the first embodiment of the present invention.

FIG. 3 is a diagram illustrating the principle of a conventional CELP encoder.

FIG. 4 is a diagram illustrating the principle of a conventional CELP decoder.

FIG. 5 shows an example of a pulse sound source.

FIG. 6 is a configuration of a pulse information codebook.

FIG. 7 is a diagram illustrating the principle of generating a pulse sound source vector.

FIG. 8 is a comparison of impulse response waveforms with and without weighting.

FIG. 9 is an explanatory diagram of a partial cross-correlation calculation method.

FIG. 10 is an explanatory diagram of a simplified convolution operation.

[Explanation of symbols]

12: linear prediction analyzer, 14: linear prediction parameter quantizer, 15, 69: linear prediction parameter quantization code book, 19: auditory weighting filter, 21: weighting synthesis filter, 25: adaptive codebook searcher, 26, 6
0 ... Adaptive codebook, 32 ... Pulse sound source searcher, 3
3, 64 pulse generator, 35, 62 pulse information codebook, 39 gain optimizer / quantizer, 40, 57
Gain codebook, 71: synthesis filter, 73: post filter.

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-5-346798 (JP, A) JP-A-5-165497 (JP, A) JP-A-4-55899 (JP, A) 101800 (JP, A) JP-A-4-107599 (JP, A) JP-A-4-58299 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 19/12

Claims (6)

(57) [Claims] An encoding unit performs a short-term prediction analysis of an input speech at predetermined time intervals (frames), and performs a long-term prediction analysis at a time interval (subframe) equal to or shorter than the frame. In a code-driven linear prediction (CELP) speech coding method for selecting an optimal code vector from a code book prepared in advance as a driving sound source for each subframe, the code vector is a pulse train having a constant amplitude and an equal interval. And selecting the optimal code vector by performing a full search of the code book. 2. The speech encoding method according to claim 1, wherein the interval between the pulse trains is a range that substantially covers a variation range of a pitch period of a human utterance. 3. The method according to claim 1, wherein the leading pulse position of the pulse train in the subframe can be set from the beginning to the end of the subframe regardless of the pulse interval. Audio coding method. 4. The selection of the optimum code vector is performed by subtracting a weighted long-term prediction vector multiplied by a gain from a signal obtained by removing the influence of the past synthesis filter from the weighted input speech, 4. The speech encoding method according to claim 1, wherein the speech encoding is performed by performing an evaluation. 5. The method according to claim 1, wherein the evaluation of the weighting error is performed based on a combination of impulse responses of a weighting synthesis filter whose length is truncated to a value equal to or less than a minimum value of the pulse train interval. The speech encoding method according to claim 4. 6. The code book according to claim 1, wherein the code book stores information on a pulse interval and a leading pulse position, and the code vector is generated from the information. Audio coding method.